botfights/frontend/src/game/tts-worker.ts

// Web Worker for Kokoro TTS — runs ONNX neural network inference off the main thread.
// This prevents the 1-5 second freezes that occur when generate() runs on the UI thread.

import type { KokoroTTS } from 'kokoro-js'

let tts: KokoroTTS | null = null

self.addEventListener('message', async (e: MessageEvent) => {
  const msg = e.data

  switch (msg.type) {
    case 'init': {
      try {
        const { KokoroTTS } = await import('kokoro-js')
        tts = await KokoroTTS.from_pretrained('onnx-community/Kokoro-82M-ONNX', {
          dtype: 'q8',
          device: null,
          progress_callback: (p) => {
            if ('progress' in p && typeof p.progress === 'number') {
              self.postMessage({ type: 'progress', progress: p.progress })
            }
          },
        })
        self.postMessage({ type: 'init-done' })
      } catch (err) {
        self.postMessage({ type: 'init-failed', error: String(err) })
      }
      break
    }

    case 'generate': {
      if (!tts) {
        self.postMessage({ type: 'generate-failed', id: msg.id, error: 'not initialized' })
        break
      }
      try {
        const result = await tts.generate(msg.text, {
          voice: msg.voice,
          speed: msg.speed,
        })
        // Copy to a standalone ArrayBuffer so we can transfer ownership (zero-copy to main thread)
        const audio = new Float32Array(result.audio)
        self.postMessage(
          { type: 'generate-done', id: msg.id, audio, sampleRate: result.sampling_rate },
          { transfer: [audio.buffer] },
        )
      } catch (err) {
        self.postMessage({ type: 'generate-failed', id: msg.id, error: String(err) })
      }
      break
    }
  }
})